The conditions for whether or not a request is allowed adding to a
completion batch are a bit hard to read, and they also have a few
issues. One is that ioerror may indeed be a random value on passthrough,
and it's being checked unconditionally of whether or not the given
request is a passthrough request or not.
Rewrite the conditions to be separate for easier reading, and only check
ioerror for non-passthrough requests. This fixes an issue with bio
unmapping on passthrough, where it fails getting added to a batch. This
both leads to suboptimal performance, and may trigger a potential
schedule-under-atomic condition for polled passthrough IO.
Fixes: f794f3351f26 ("block: add support for blk_mq_end_request_batch()")
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index a0a9007cc1e3..cc1cb5de8fb6 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -861,12 +861,22 @@ static inline bool blk_mq_add_to_batch(struct request *req,
void (*complete)(struct io_comp_batch *))
{
/*
- * blk_mq_end_request_batch() can't end request allocated from
- * sched tags
+ * Check various conditions that exclude batch processing:
+ * 1) No batch container
+ * 2) Has scheduler data attached
+ * 3) Not a passthrough request and end_io set
+ * 4) Not a passthrough request and an ioerror
*/
- if (!iob || (req->rq_flags & RQF_SCHED_TAGS) || ioerror ||
- (req->end_io && !blk_rq_is_passthrough(req)))
+ if (!iob)
return false;
+ if (req->rq_flags & RQF_SCHED_TAGS)
+ return false;
+ if (!blk_rq_is_passthrough(req)) {
+ if (req->end_io)
+ return false;
+ if (ioerror < 0)
+ return false;
+ }
if (!iob->complete)
iob->complete = complete;
--
Jens Axboe
